Creative systems degrade quietly. A model version changes, a prompt is edited, a connector returns a slightly different shape, and the output gets marginally worse in a way nobody notices until a client does.
An eval is the boring fix: twenty representative inputs, an agreed view of what good looks like, run automatically on every change. It does not need to be sophisticated to be the difference between catching a regression on Tuesday and hearing about it in a review.
For creative work the scoring is usually a person comparing outputs side by side, which is fine. The value is in the fixed inputs and the regular cadence, not in automating the judgement.
How do you test a creative workflow?
A fixed set of representative inputs, an agreed view of good, and a side-by-side comparison on every change. The rigour is in the fixed inputs, not in automating taste.
Why do creative systems degrade without anyone noticing?
Because the inputs vary constantly, so a small drop in quality is indistinguishable from a hard brief. A fixed eval set removes that excuse.